A method for enhancing video monitoring image quality under adverse weather conditions

Abstract

The present application relates to a kind of video monitoring image quality enhancement methods under adverse weather conditions, comprising the following steps: construct the blur rendering model based on generative adversarial network, then train and verify model;Blur rendering model includes discriminator, residual dense network, feature fusion module and generator;Initial image input blur rendering model is first judged by discriminator whether real sample or generated image, then, initial image input residual dense network, feature information of input image is extracted step by step, foreground target feature is strengthened, obtain multiple feature maps, generate feature fusion map after the fusion of multiple feature maps, generator adds random noise to feature fusion map, output generated image;Image denoising restoration model based on Unet model is constructed, and model is trained and verified after;Generated image is input into image denoising restoration model, and clear restored image is output.The present application has the advantages that foreground target information can be enhanced, and the authenticity and accuracy of image recovery can be ensured.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence image processing technology, and in particular to a method for enhancing the quality of video surveillance images under adverse weather conditions. Background Technology

[0002] Digital safety management at power work sites primarily relies on video surveillance equipment to capture on-site footage. After frame extraction and processing, artificial intelligence algorithms are used for identification and analysis. However, due to the prevalence of outdoor work sites in power construction, the video images captured by monitoring equipment are often of poor quality and blurry in adverse weather conditions such as rain, fog, and snow. This severely impacts the effectiveness of video surveillance, causing a sharp decline in the accuracy of identifying risky behaviors based on visual recognition technology. This becomes a significant obstacle to the viewing experience and the accuracy of target and abnormal behavior identification, greatly hindering remote supervision and digital safety management. Therefore, improving the quality of video surveillance images under adverse weather and lighting conditions, and enhancing the recognizability of on-site video footage, is an urgent issue that needs to be addressed.

[0003] A key challenge in video surveillance image quality enhancement technology is maintaining image quality under asymmetric encoding and decoding conditions. Asymmetric video encoding and decoding refers to an uneven distribution of processing power and time during compression and decompression. Compression requires significant processing power and time, while decompression allows for better real-time playback; compression and decompression occur at different speeds. This can easily lead to video data corruption and image quality issues. Therefore, it is necessary to employ different algorithmic strategies or technical parameters during video compression and decompression to optimize encoding efficiency and decoding quality, address data corruption and image quality problems during encoding and decoding, and improve the quality of video surveillance at power operation sites under adverse weather conditions. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method for enhancing the quality of video surveillance images under adverse weather conditions, which can enhance foreground target information, improve image quality, and ensure the authenticity and accuracy of image restoration.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for enhancing the quality of video surveillance images under adverse weather conditions includes the following steps: constructing a bokeh rendering model based on a generative adversarial network, and then training and validating the model; the bokeh rendering model includes a discriminator, a residual dense network, a feature fusion module, and a generator; the initial image input to the bokeh rendering model is first determined by the discriminator to determine whether the initial image is a real sample or a generated image output by the generator; the initial image is input to the residual dense network, which extracts feature information of the input image step by step, enhances foreground target features, and obtains multiple feature maps; then the feature fusion module fuses the multiple feature maps and inputs them into the generator; the generator takes random noise as input and outputs a generated image; constructing an image denoising and restoration model based on the Unet model, and then training and validating the model; the generated image is input into the image denoising and restoration model, and a clear restored image is output.

[0007] More preferably, the discriminator employs a downsampling method, and its network structure has five layers, including four convolutional layers and one fully connected layer. The convolution result of each convolutional layer is normalized before being input into the next layer. The fully connected layer serves as the output layer, outputting the discrimination result. The results of each convolutional layer are converged using the LeakyReLU() activation function, and the fully connected layer outputs a probability using the sigmoid() activation function, representing the authenticity of the initial image.

[0008] More preferably, the residual dense network includes multiple cascaded improved residual modules, each residual module including multiple convolutional layers, each convolutional layer adopting a dense connection structure, and each convolutional layer is connected to all its subsequent convolutional layers in a feedforward manner, so that the inter-layer connections are tight and the features of the previous convolutional layers are fully utilized, the transfer of inter-layer features is enhanced, and multiple feature maps are output.

[0009] More preferably, the feature fusion module first performs global average pooling on the input feature maps of each layer, then learns the correlation between channels through a fully connected layer, then obtains the weights of the corresponding channels of different layers of features through splicing and slicing and the softmax function, and finally multiplies the weights of each layer of features with the corresponding layer of feature maps input to the feature fusion module, and then superimposes and fuses the results to obtain the feature fusion map.

[0010] More preferably, the generator generates feature images of different sizes through an upsampling method, inputs a feature fusion map and noise, sets the corresponding convolution kernel according to the required feature map size, extracts feature information from the input image step by step, and then continuously optimizes the details of the image according to the depth of the network, and outputs the generated image.

[0011] More preferably, the image denoising and restoration model is a Unet model with added noise removal network modules (DNB). Specifically, the Unet model is a fully convolutional network (FCN) consisting of an encoder and a decoder connected by skip connections. The input generated image is compressed by an encoder composed of convolutional and pooling layers to form an implicit data representation. Then, the image is enlarged by a decoder composed of convolutional and transposed convolutional layers, and finally the restored image is output. Multiple noise removal network modules (DNB) are introduced into the encoder and decoder. The noise removal network modules (DNB) consist of three parts: the first part is the first layer, which consists of 3*3 convolutional layers and ReLU activation functions; the second part consists of layers 2 to 16, which consist of 15 unit blocks; and the third part consists of 3*3 convolutional layers.

[0012] The present invention has the following beneficial effects:

[0013] 1. This invention utilizes Generative Adversarial Network (BGGAN)-based bokeh rendering technology, implemented through an end-to-end convolutional neural network generator and a multi-receptive-field discriminator. The generator outputs the bokeh-rendered image, while the discriminator monitors the difference between the generated image and the real bokeh image. This approach enhances the foreground target information in images captured by surveillance cameras in power grid application scenarios, thereby improving image quality. Furthermore, during training, PatchGAN discriminators of varying depths are employed to form a multi-receptive-field discriminator that focuses on details at different image sizes, further enhancing the naturalness and realism of the bokeh effect.

[0014] 2. This invention is an image enhancement technique based on denoising principles. It employs an improved UNet architecture for image restoration, aiming to recover high-quality images from low-quality input images. It is particularly suitable for images heavily influenced by environmental factors. Through a hierarchical encoder-decoder structure, multi-scale features are extracted, and skip connections are used to reconstruct image details, achieving efficient noise removal and image quality enhancement. Furthermore, the introduction of a noise removal network block (DNB) and an attention mechanism further enhances the model's ability to identify and remove noise, ensuring the realism and accuracy of the restored image. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the image quality enhancement process of the present invention;

[0016] Figure 2 This is a schematic diagram of the generative adversarial network model of the present invention;

[0017] Figure 3 This is a diagram of the generator code structure of the present invention;

[0018] Figure 4 This is a schematic diagram of the dense connection structure of the present invention;

[0019] Figure 5 This is a schematic diagram of the feature fusion process of the present invention;

[0020] Figure 6 This is a schematic diagram of the architecture of the Unet model of the present invention;

[0021] Figure 7 This is a schematic diagram of the denoising algorithm framework of the present invention;

[0022] Figure 8 This is a structural diagram of the Unitblock of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0024] A method for enhancing the quality of video surveillance images under adverse weather conditions includes the following steps:

[0025] Step 10, please refer to Figure 1 and Figure 2 We constructed a bokeh rendering model based on generative adversarial networks, and then trained and validated the model.

[0026] The bokeh rendering model includes a discriminator, a residual dense network, a feature fusion module, and a generator. First, the initial image input to the bokeh rendering model is determined by the discriminator to be either a real sample or a generated image output by the generator. Second, the initial image is input to the residual dense network, which extracts feature information from the input image step by step to obtain multiple feature maps. Then, the feature fusion module fuses the multiple feature maps and inputs them into the generator. The generator takes random noise as input and outputs the generated image.

[0027] The discriminator employs a downsampling method and has a five-layer network structure. The entire network structure lacks pooling layers, including four convolutional layers and one fully connected layer. The convolution result of each convolutional layer is normalized before being input to the next layer. The fully connected layer serves as the output layer, outputting the discrimination result, ranging from 0 to 1, indicating whether the input image belongs to a real sample or is an image generated by the generator. The results of each convolutional layer are converged using the LeakyReLU() activation function, and the fully connected layer outputs a probability using the sigmoid() activation function, representing the authenticity of the initial image. A batch normalization (BN) layer is also added to the discriminator, and each layer uses an activation function for non-linear processing. The specific hierarchical structure is shown in Table 1 below: DCGAN Discriminator Model Hierarchical Structure.

[0028] Table 1:

[0029]

[0030]

[0031] The specific explanation for Table 1 is as follows:

[0032] Layer 1: Input layer. First, an image is input, with Stride set to 1 and Padding set to 0. Convolution is performed using the Conv2d() function, followed by the LeakyReLU() activation function to optimize network convergence speed.

[0033] Layer 2: Input image, Stride set to 2, Padding set to 1, add BN layer to batch normalize the data, and use LeakyReLU() activation function to accelerate convergence;

[0034] Layer 3: Input image, Stride set to 2, Padding set to 1, add BN layer to batch normalize the data, and use LeakyReLU() activation function to accelerate convergence;

[0035] Layer 4: Input image, Stride set to 2, Padding set to 1, add BN layer to batch normalize the data, and use LeakyReLU() activation function to accelerate convergence;

[0036] Layer 5: Output layer, Stride is set to 3, Padding is set to 1, and the sigmoid() function is used to output probabilities, representing the realism of the image.

[0037] The generator produces feature images of different sizes using an upsampling method. The network structure has five layers and no pooling layers. The input is a fused feature map and noise of dimension nz, which follows a uniform distribution. The appropriate convolutional kernel is set according to the required feature map size. The size of the convolutional kernel affects the size of the learnable feature values ​​in the generator network. Then, feature information such as lines and style is extracted step-by-step from the input image. Subsequently, based on the network depth, the image details are continuously optimized, and the generated image is output.

[0038] The `nn.Sequential()` function in the code adds convolutions, activations, etc., to the module in the correct order based on the input sequence. After the network layers are built, the `forward` method is automatically called. Batch normalization (BN) is added to each layer except the output layer to mitigate model crashes. The corresponding activation function is used according to the needs of each network layer. Finally, a 3-channel RGB image of a single pixel is output. The generator code structure diagram is shown below. Figure 3 As shown in Table 2, the specific hierarchical structure of the DCGAN generative model is as follows:

[0039] Table 2:

[0040]

[0041] The specific explanation for Table 2 is as follows:

[0042] Layer 1 is the input layer, which takes random noise as input, with the Stride set to 1 and the Padding set to 0. Next, the ConvTranspose2d() function is used to perform inverse convolution, and the data is batch normalized in the BN layer. The ReLU activation function is used to accelerate convergence, and the corresponding feature map is output.

[0043] Layer 2: Set the Stride to 2 and the Padding to 1, add a BN layer to batch normalize the data, use the ReLU activation function to accelerate convergence, and output the corresponding feature map.

[0044] Layer 3: Set the Stride to 2 and the Padding to 1, add a BN layer to batch normalize the data, use the ReLU activation function to accelerate convergence, and output the corresponding feature map.

[0045] Layer 4: Set the Stride to 2 and the Padding to 1, add a BN layer to batch normalize the data, use the ReLU activation function to accelerate convergence, and output the corresponding feature map.

[0046] Layer 5: Output layer, with Stride set to 3, Padding set to 1, and Tanh used as the activation function to output the corresponding image.

[0047] In summary, it can be seen that the settings for Padding, Stride, etc., of the generator and discriminator network models are almost identical, and the overall network structure is symmetrical, with only the activation functions used in each layer being different.

[0048] It is worth noting that the generator and discriminator are trained and optimized alternately.

[0049] The discriminator classifies real data from fake data generated by the generator. The discriminator's loss function penalizes misclassifications, such as classifying a real instance as fake or a fake instance as real. The discriminator calculates its loss and updates its weights using backpropagation, employing an error backpropagation mechanism. During discriminator training, the generator is not trained; that is, when the generator generates example data for the discriminator, its weights remain constant. In PyTorch, this is achieved by using `detach()` to prevent the generator's gradient from being calculated.

[0050] For the generator, random noise samples are used as input; the generator generates output from the sampled random noise samples, and the discriminator judges whether the output is "true" or "false", which is used as the generator's output; the error loss is calculated from the classification output of the discriminator; the gradient is obtained through backpropagation between the discriminator and the generator; the gradient is used to update the generator's weights.

[0051] Alternating training process: During the discriminator training phase, the generator remains unchanged. Because discriminator training attempts to distinguish real data from fake data, the discriminator must learn how to identify the generator's flaws. This is the difference between a fully trained generator and an untrained generator that can only produce random outputs. Similarly, during the generator training phase, the discriminator remains unchanged. Otherwise, the generator, like trying to hit a moving target, might never converge.

[0052] To better preserve and utilize the features extracted from each layer, dense connections and feature fusion are introduced on the basis of residual networks, and a new image super-resolution reconstruction algorithm using residual dense networks is proposed. Dense connections are employed in the feature extraction part of the network to enhance feature transfer between layers. The proposed feature fusion module merges inter-layer features to reduce feature dimensionality, achieving efficient and full utilization of feature information from each layer.

[0053] The residual dense network comprises multiple cascaded improved residual modules, such as... Figure 4 As shown, each incomplete module includes multiple convolutional layers. Each convolutional layer adopts a densely connected structure, and each convolutional layer is connected to all subsequent convolutional layers using a feedforward method. This ensures tight inter-layer connections, fully utilizes the features of previous convolutional layers, enhances the transfer of features between layers, and outputs multiple feature maps. Figure 4 As shown, in the densely connected structure, the first... i The input to the layer includes feature maps F0, F1, ..., F from the outputs of all previous layers. i-1 The feature map F output by this layer i It will also become the input to all subsequent layers. The layer contains a convolutional layer and a non-linear activation function, the first... i Feature map F output by layer i This can be represented as: fi = g(wi*[F0, F1, ..., F i-1 ]+b i In the formula, wi and b i Let F0, F1, ..., F1 represent the weights and offsets of the i-th layer, respectively. i-1The symbol ] represents the concatenation of feature maps, and g is the activation function. Dense connections connect the features of all convolutional layers in a concatenated manner, where the feature maps of each layer are passed to all subsequent layers. This improves reconstruction performance by preserving and utilizing features from each layer without information loss. This not only improves the vanishing gradient problem but also enhances the transfer of features between layers. Dense connections achieve feature reuse by concatenating features on channels, resulting in better reconstruction performance than residual networks.

[0054] In residual dense networks, the information in the feature maps output by different convolutional layers complements each other, resulting in richer and more effective inter-layer feature information. Furthermore, dense connections concatenate feature maps from different layers as input to the subsequent upsampling part; if these feature maps were directly fed into the upsampling part, the computational cost would increase significantly. To more efficiently and fully utilize the feature information transmitted through dense connections, a feature fusion module is proposed.

[0055] Please see Figure 5 First, the feature fusion module performs global average pooling on the input feature maps of each layer, which reduces the feature dimension, increases the receptive field, and removes redundant information to improve the network computation efficiency.

[0056] Next, the correlation between channels is learned through the fully connected layer. Let the number of feature maps in the input feature fusion module be N, and the number of channels in the feature maps be C. The output of the fully connected layer is represented as follows:

[0057] Yi = a i X i ;

[0058] In the formula, i = 1, 2, ..., N, X i =[x i1 x i2 , ..., x iC ], Yi = [y i1 y i2 , ..., y iC ], X i M represents the i-th feature map of the input feature fusion module. i The global feature obtained after global average pooling, a i X represents i The corresponding weights of the fully connected layer.

[0059] Then, the weights of the corresponding channels of the features at different layers are obtained by concatenating slices and using the softmax function. The weights of each channel of the features at different layers are represented as follows:

[0060] W j =F S (Y j )

[0061] In the formula, j = 1, 2, ..., C, Y j =[y 1j y 2j , ..., y Nj ], W j =[w 1j w 2j ,…,w Nj ], F S This represents the softmax function.

[0062] Finally, the feature weights of each layer are multiplied by the corresponding layer feature maps of the input feature fusion module, and the results are superimposed and fused to obtain the feature fusion map. The output of the feature fusion module is R = [r1, r2, ..., r...]. c It consists of the following:

[0063]

[0064] In the formula, m ij The i-th feature map M is input to the feature fusion module. i The j-th channel.

[0065] The feature fusion module merges inter-layer features to reduce feature dimensionality, efficiently and fully utilizing feature information from each layer, which helps improve network computational efficiency and super-resolution reconstruction results.

[0066] Step 20: Construct an image denoising and restoration model based on the Unet model, then train and validate the model, input the generated image into the image denoising and restoration model, and output a clear restored image.

[0067] The image denoising and restoration model described is the Unet model with the addition of a noise removal network module, DNB. Please refer to [link / reference]. Figure 6 The Unet model is a fully convolutional network (FCN) consisting of an encoder and a decoder connected by skip connections. The input generated image is compressed by an encoder composed of convolutional and pooling layers to form an implicit data representation. Then, the image is enlarged by a decoder composed of convolutional and transposed convolutional layers, and finally the restored image is output.

[0068] The encoder's primary task is to progressively compress the spatial resolution of the input image, extracting higher-level features. This part contains a series of convolutional layers and max pooling layers, with each pooling operation reducing the image's spatial dimensions by half. The decoder's task is to progressively restore the image's spatial resolution, mapping the high-level features extracted by the encoder back to the original image resolution. The decoder includes deconvolution (upsampling) operations and combines them with corresponding feature layers from the encoder to achieve fine-grained boundary restoration. The output feature map of each encoder layer is concatenated with the feature map of the corresponding layer in the decoder, forming skip connections. This allows for the fusion of local information from the encoder and global information from the decoder, thereby improving the accuracy of the segmentation results.

[0069] In the encoder and decoder, multiple noise removal network modules (DNBs) are introduced. These DNBs focus on noise recognition and removal by incorporating components such as layer normalization, convolutional operations of various sizes, and attention mechanisms. Through fine-grained feature processing, DNBs can effectively reduce image noise while preserving important image details.

[0070] Layer normalization: This normalizes the activation values ​​of all neurons in each layer of the neural network, converting the activation values ​​of each layer into a distribution with a mean of 0 and a standard deviation of 1, and then scaling and shifting the results. Given the input vector Z = (Z1, Z2, ..., Zn) of a certain layer in the neural network... H ), where H is the number of neurons in this layer, calculated using the following formula:

[0071]

[0072] In the formula, μ is the mean of the activation values ​​of all neurons in the layer, and o is the standard deviation of the activation values ​​of all neurons in the layer.

[0073]

[0074] In the formula, ∈ is a small positive number used to prevent division by zero errors. After normalization, each normalized activation value Zi undergoes scaling and offset operations to maintain the model's expressive power. This process is accomplished by introducing learnable parameters γ and β:

[0075] Y i =γZ i +β;

[0076] This ensures that the input to each layer remains stable during training, which helps improve the training efficiency and stability of the model.

[0077] Convolutions of various sizes: Advanced denoising algorithm frameworks in deep learning, such as Figure 7As shown, it can be mainly divided into three parts. The first part is the first layer, which consists of a 3×3×64 convolutional layer and a ReLU activation function. The second part consists of 15 unit blocks from layers 2 to 16. The unit blocks are as follows: Figure 8 As shown, the network consists of 3×3×64 convolutional layers, Batch Normalization (BN) layers, and the ReLU activation function; the third part consists of 3×3 convolutional layers. The entire network performs a global jump from the input to the output, forming a residual learning strategy. The entire network can be described as follows:

[0078] G(x,y)=O(x,y)+b(x,y);

[0079] In the formula: G(x, y) is a two-dimensional noisy image; O(x, y) is the original clean image without noise; b(x, y) represents the added noise. A noisy image is a degraded result of the original image under additive noise. Therefore, the goal of image denoising is to obtain a ^O(x, y), which is an estimate of the original image. The smaller the difference between O(x, y) and ^O(x, y), the better the image denoising effect.

[0080] To further improve the quality of image restoration, this scheme also incorporates attention mechanisms into DNB, such as distraction attention and skip attention modules. These attention modules help the network focus more on key areas of the image, optimize the information flow during image reconstruction, and thus preserve more details and texture information while denoising.

[0081] Step 1: Preprocess the training dataset by using Gaussian white noise to simulate real noise and adding it to the original image to obtain the corresponding noise image x_noise_img;

[0082] Step 2: Input the noisy image xnoise_img into the denoising network, and the output is the noise distribution image xno1se;

[0083] Step 3: Subtract the noise distribution map xnoise from the noisy image x_noise_img to obtain the denoising result x_dn using the denoising network;

[0084] Step 4: Input the noisy image xnoise_img into the attention mechanism network, and the output is the noise weight image noise mask;

[0085] Step 5: Combine the noise distribution image xnoise output from Step 2 and the noise weight image noise_mask output from Step 4 to obtain the noise residual image noise.

[0086] Step 6: Subtract the noise residual image noise from the noisy image xnoise_img to obtain the denoised clean image xmask dn;

[0087] Step 7: Based on xdn obtained in Step 3 and xmaskdn obtained in Step 6, the mean squared error (MSE) function is used as the loss function to continuously learn the mapping relationship between noisy images and clean images, and adjust the network model parameters until the network converges.

[0088] This invention innovatively proposes a bokeh rendering technique based on generative adversarial networks (BGGAN). This technique utilizes an end-to-end convolutional neural network generator and a multi-receptive-field discriminator to achieve image quality processing, addressing the problem of maintaining video surveillance quality under asymmetric encoding and decoding conditions. Simultaneously, an image enhancement technique based on image denoising principles employs the UNet architecture and its improvements for image restoration, solving the problem of low efficiency in noise removal and image quality improvement.

[0089] The above description is merely a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for enhancing video surveillance image quality under adverse weather conditions, characterized in that: It comprises the following steps: A bokeh rendering model based on a generative adversarial network is constructed, and then the model is trained and verified; the bokeh rendering model comprises a discriminator, a residual dense network, a feature fusion module and a generator; an initial image input into the bokeh rendering model is first judged by the discriminator to be a real sample or a generated image output by the generator, then the initial image is input into the residual dense network, feature information of the input image is extracted step by step, foreground target features are strengthened, a plurality of feature maps are obtained, the feature fusion module fuses the plurality of feature maps to generate a feature fusion map, and the feature fusion map is input into the generator, the generator adds random noise to the feature fusion map, and outputs a generated image; An image denoising and restoration model based on a Unet model is constructed, and then the model is trained and verified; the generated image is input into the image denoising and restoration model, and a clear restored image is output. The image denoising and restoration model adds a noise removal network module DNB in the Unet model, specifically, the Unet model is a full convolutional network FCN, which comprises an encoder and a decoder and is connected through a jump connection; the input generated image is compressed by the encoder composed of a convolutional layer and a pooling layer to form a data implicit representation, and then the image is enlarged by the decoder composed of a convolutional layer and a transposed convolutional layer, and finally the restored image is output; in the encoder and the decoder, a plurality of noise removal network modules DNB are introduced, specifically, the noise removal network module DNB comprises layer normalization, convolution operations of multiple sizes and an attention mechanism; Layer normalization: it normalizes all neuron activation values of each layer of the neural network, converts the activation values of each layer into a distribution with a mean of 0 and a standard deviation of 1, and then scales and shifts the result; given an input vector of a certain layer in the neural network wherein H is the number of neurons of the layer, and the calculation formula is as follows: ； ； where μ is the mean of all neuron activation values of the layer, is the standard deviation of all neuron activation values of the layer; ； In the formula, ε is a small positive number, which is used to prevent division by zero error; After normalization, each normalized activation value Scaling and shifting operations are performed to preserve the expressiveness of the model; this is done by introducing learnable parameters γ and β: ； In this way, it can be ensured that the input of each layer remains stable during the training process, which helps to improve the training efficiency and stability of the model; The convolution operations of multiple sizes are divided into three parts, the first part is the first layer composed of a convolutional layer 3x3x64 and a ReLU activation function; the second part is 15 Unit blocks from the 2nd to the 16th layer, each Unit block is composed of a convolutional layer 3x3x64, a BN layer and an activation function ReLU; the third part is composed of a convolutional layer 3x3; the entire network performs global jumping from the input end to the output end, forming a residual learning strategy, and the entire network is described as: ； where: is a two-dimensional noise image; is a noise-free original clean image; represents additional noise; the purpose of image denoising is to obtain a , is an estimate of the original image; and The smaller the difference between and, the better the image denoising effect.

2. The method for enhancing the image quality of video monitoring in adverse weather conditions according to claim 1, characterized in that: The discriminator adopts a downsampling method, and its network structure has five layers, including four convolutional layers and one fully connected layer, and the convolutional results of each convolutional layer are input into the next layer after normalization processing, the fully connected layer is used as an output layer to output a discrimination result, the results of each convolutional layer are converged by using a LeakyReLU() activation function, and the fully connected layer outputs a probability by using a sigmoid() activation function, which represents the authenticity of the initial image.

3. The method for enhancing the image quality of video monitoring in adverse weather conditions according to claim 1, characterized in that: The residual dense network comprises a plurality of improved residual modules connected in series, each residual module comprises a plurality of convolutional layers, each convolutional layer adopts a dense connection structure, and each convolutional layer is connected with all subsequent convolutional layers in a feedforward manner, so that the layers are closely connected and the features of the previous convolutional layers are fully utilized, the transmission of interlayer features is enhanced, and a plurality of feature maps are output.

4. The method for enhancing the image quality of video monitoring in adverse weather conditions according to claim 1, characterized in that: The feature fusion module firstly performs global average pooling on each layer of the input feature map respectively, then learns the correlation between channels through a fully connected layer, then obtains the weight of the corresponding channel of different layers of features through splicing and a softmax function, and finally multiplies each layer of feature weight and the corresponding layer of feature map of the input feature fusion module, and stacks and fuses the results to obtain a feature fusion map.

5. The method of claim 4, wherein the method further comprises: The generator generates feature images of different sizes through an upsampling method, inputs the feature fusion map and noise, sets a corresponding convolution kernel according to the required feature map size, extracts feature information of the input image in steps, then constantly optimizes the details of the picture according to the depth of the network, and outputs the generated image.

Images